新文件： doc/tutorial/onnx-op/dynamo-export.ipynb

	新文件：   doc/tutorial/onnx-op/export.ipynb

doc/tutorial/.gitignore

@@ -2,3 +2,4 @@ data/
 params/
 attack/
 *.pt
+*.sarif

doc/tutorial/onnx-op/dynamo-export.ipynb

@@ -0,0 +1,256 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# PyTorch 转 ONNX \n",
+    "\n",
+    "参考：[PyTorch 转换为 ONNX](https://pytorch.org/tutorials//beginner/onnx/export_simple_model_to_onnx_tutorial.html)\n",
+    "\n",
+    "在 PyTorch 2.1 版本中，有两种 ONNX 导出工具。\n",
+    "\n",
+    "- {func}`torch.onnx.dynamo_export` 是最新的（仍处于测试阶段）基于 TorchDynamo 技术的导出器，该技术与 PyTorch 2.0 一同发布。\n",
+    "- {func}`torch.onnx.export` 是基于 TorchScript 后端的，自 PyTorch 1.2.0 以来一直可用。\n",
+    "\n",
+    "由于 ONNX 导出器使用 `onnx` 和 `onnxscript` 将 PyTorch 算子转换为 ONNX 算子，需要安装：\n",
+    "\n",
+    "```bash\n",
+    "pip install onnx onnxscript\n",
+    "```\n",
+    "\n",
+    "下面以简单的分类器为例展开。\n",
+    "\n",
+    "## 简单的分类器模型导出准备"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "import torch.nn as nn\n",
+    "import torch.nn.functional as F\n",
+    "\n",
+    "\n",
+    "class MyModel(nn.Module):\n",
+    "\n",
+    "    def __init__(self):\n",
+    "        super(MyModel, self).__init__()\n",
+    "        self.conv1 = nn.Conv2d(1, 6, 5)\n",
+    "        self.conv2 = nn.Conv2d(6, 16, 5)\n",
+    "        self.fc1 = nn.Linear(16 * 5 * 5, 120)\n",
+    "        self.fc2 = nn.Linear(120, 84)\n",
+    "        self.fc3 = nn.Linear(84, 10)\n",
+    "\n",
+    "    def forward(self, x):\n",
+    "        x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2))\n",
+    "        x = F.max_pool2d(F.relu(self.conv2(x)), 2)\n",
+    "        x = torch.flatten(x, 1)\n",
+    "        x = F.relu(self.fc1(x))\n",
+    "        x = F.relu(self.fc2(x))\n",
+    "        x = self.fc3(x)\n",
+    "        return x"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 将模型导出为 ONNX 格式\n",
+    "\n",
+    "实例化模型并创建随机的 32x32 输入。接下来，可以将模型导出为 ONNX 格式。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "/media/pc/data/tmp/cache/conda/envs/py311/lib/python3.11/site-packages/torch/onnx/_internal/exporter.py:130: UserWarning: torch.onnx.dynamo_export only implements opset version 18 for now. If you need to use a different opset version, please register them with register_custom_op.\n",
+      "  warnings.warn(\n"
+     ]
+    }
+   ],
+   "source": [
+    "torch_model = MyModel()\n",
+    "torch_input = torch.randn(1, 1, 32, 32)\n",
+    "onnx_program = torch.onnx.dynamo_export(torch_model, torch_input)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "不需要对模型进行任何代码更改。生成的 ONNX 模型存储在二进制 protobuf 文件 `torch.onnx.ONNXProgram` 中。"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 将 ONNX 模型保存到文件中\n",
+    "\n",
+    "尽管在许多应用中将导出的模型加载到内存中是有用的，但我们可以将其保存到磁盘上，代码如下："
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "onnx_program.save(\"my_image_classifier.onnx\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "您可以将 ONNX 文件重新加载到内存中，并使用以下代码检查其格式是否正确："
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import onnx\n",
+    "onnx_model = onnx.load(\"my_image_classifier.onnx\")\n",
+    "onnx.checker.check_model(onnx_model)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 使用 ONNX Runtime 执行 ONNX 模型\n",
+    "\n",
+    "最后一步是使用 ONNX Runtime 执行 ONNX 模型，但在我们这样做之前，让我们先安装 ONNX Runtime。\n",
+    "\n",
+    "```bash\n",
+    "pip install onnxruntime\n",
+    "```\n",
+    "\n",
+    "ONNX 标准不支持 PyTorch 支持的所有数据结构和类型，所以需要在将输入喂给 ONNX Runtime 之前，将 PyTorch 的输入适配为 ONNX 格式。在我们的示例中，输入恰好是相同的，但在更复杂的模型中，它可能比原始的 PyTorch 模型有更多的输入。\n",
+    "\n",
+    "ONNX Runtime 需要额外的步骤，该步骤涉及将所有 PyTorch 张量转换为 Numpy（在 CPU 上），并在字典中包装它们，其中键是字符串，表示输入名称，值为 `numpy` 张量。\n",
+    "\n",
+    "现在我们可以创建 ONNX Runtime 推理会话，使用处理过的输入执行 ONNX 模型并获取输出。在这个教程中，ONNX Runtime 是在 CPU 上执行的，但它也可以在 GPU上 执行。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Input length: 1\n",
+      "Sample input: (tensor([[[[-0.5305, -0.6818,  2.2350,  ..., -0.2503,  0.4694,  1.3666],\n",
+      "          [ 0.7013,  0.0179, -1.2689,  ...,  0.4369,  0.5982, -0.6541],\n",
+      "          [ 0.8644,  0.8552,  0.4100,  ..., -0.8513,  0.4207,  0.4363],\n",
+      "          ...,\n",
+      "          [ 0.4400, -0.3064, -1.9848,  ...,  0.0462,  0.7269,  1.3543],\n",
+      "          [ 1.5511, -0.6354,  0.9151,  ...,  0.2501, -0.0140, -0.3875],\n",
+      "          [-1.2229, -0.8693,  1.0505,  ...,  0.0598,  0.7852,  0.1350]]]]),)\n"
+     ]
+    }
+   ],
+   "source": [
+    "import onnxruntime\n",
+    "\n",
+    "onnx_input = onnx_program.adapt_torch_inputs_to_onnx(torch_input)\n",
+    "print(f\"Input length: {len(onnx_input)}\")\n",
+    "print(f\"Sample input: {onnx_input}\")\n",
+    "\n",
+    "ort_session = onnxruntime.InferenceSession(\"./my_image_classifier.onnx\", providers=['CPUExecutionProvider'])\n",
+    "\n",
+    "def to_numpy(tensor):\n",
+    "    return tensor.detach().cpu().numpy() if tensor.requires_grad else tensor.cpu().numpy()\n",
+    "\n",
+    "onnxruntime_input = {k.name: to_numpy(v) for k, v in zip(ort_session.get_inputs(), onnx_input)}\n",
+    "\n",
+    "onnxruntime_outputs = ort_session.run(None, onnxruntime_input)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 将PyTorch的结果与ONNX Runtime的结果进行比较\n",
+    "\n",
+    "确定导出模型是否良好的最佳方式是通过与 PyTorch 的数值评估，这是我们的真实来源。\n",
+    "\n",
+    "为此，我们需要使用相同的输入执行 PyTorch 模型，并将结果与 ONNX Runtime 的结果进行比较。在比较结果之前，我们需要将 PyTorch 的输出转换为匹配 ONNX 的格式。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "PyTorch and ONNX Runtime output matched!\n",
+      "Output length: 1\n",
+      "Sample output: [array([[-0.02100155, -0.13608684, -0.14742026, -0.04622332, -0.01618233,\n",
+      "        -0.07353653, -0.11702952, -0.02780916,  0.09021657,  0.02800114]],\n",
+      "      dtype=float32)]\n"
+     ]
+    }
+   ],
+   "source": [
+    "torch_outputs = torch_model(torch_input)\n",
+    "torch_outputs = onnx_program.adapt_torch_outputs_to_onnx(torch_outputs)\n",
+    "\n",
+    "assert len(torch_outputs) == len(onnxruntime_outputs)\n",
+    "for torch_output, onnxruntime_output in zip(torch_outputs, onnxruntime_outputs):\n",
+    "    torch.testing.assert_close(torch_output, torch.tensor(onnxruntime_output))\n",
+    "\n",
+    "print(\"PyTorch and ONNX Runtime output matched!\")\n",
+    "print(f\"Output length: {len(onnxruntime_outputs)}\")\n",
+    "print(f\"Sample output: {onnxruntime_outputs}\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "xin",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.11.8"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}